基于深度学习的粘虫板储粮害虫图像检测算法的研究

本文提出了一种基于深度学习的粘虫板储粮害虫图像检测算法，实现了放置在粮仓表面粘虫板诱捕的六大类害虫（米象/玉米象、谷蠹、扁谷盗、锯谷盗、拟谷盗、烟草甲）的定位和识别。考虑粘虫板图像的背景复杂、害虫体积较小、姿态多样的特点，改进了SSD的目标框回归策略、损失函数和特征提取网络结构，测试结果表明本文提出的算法可有效检测粘虫板上的害虫，检测平均正确率(mAP)可以达到81.36%。改进后的SSD算法部署在储粮害虫监测系统中，目前已在全国十一个粮库进行实验测试。

Research on detection of stored-grain insects image on sticky board using deep learning

A detection algorithm for stored-grain insects from image on sticky boards was developed by applying deep learning. This algorithm is mainly aimed at the following six species of common stored-grain insects: Cryptoleste Pusillus(S.), Sitophilus Oryzae(L.), Oryzaephilus Surinamensis(L.), Tribolium Confusum(Jaquelin Du Val), Rhizopertha Dominica(F.). The images of sticky boards have the characteristics of complex background, small size of insect and various postures. Based on these analyses, we improved the target box regression strategy, loss function and feature extraction network structure of SSD. The evaluation results show that the algorithm we developed can effectively detect the stored-grain insects on sticky boards, and the mean average precision (mAP) is 81%. The improved SSD algorithm has been applied to the stored grain insects monitoring system developed by our group, and has been placed in 11 grain depots in China.